Prompt injection in offshore FPSO gas compression AI

Four adversarial injection surfaces in offshore FPSO gas compression AI

1. Gas compressor anti-surge control display AI (Compressor Controls Corporation CCC AI, Atlas Copco Comptrol AI, Siemens Dresser-Rand DATUM AI — FPSO centrifugal gas compressor surge control display AI)

Compressor surge — the condition in which a centrifugal compressor operating point moves to the left of the surge line on the compressor map (low flow, high differential pressure), causing flow reversal through the compressor stages, high-energy oscillatory pressure waves at the compressor discharge, and rapid mechanical loading of the impeller, diffuser, and shaft — is the most severe operational hazard for centrifugal gas compressors in FPSO gas compression service. A single severe surge cycle in an FPSO associated gas compressor (handling gas at 100–300 bar discharge pressure, 5–30 MW motor or gas turbine drive) can: fracture compressor impellers (aluminium or titanium alloy, operating at tip speeds of 300–500 m/s); damage labyrinth seals and dry gas seals (DGS), releasing high-pressure associated gas into the compressor hall; and initiate a flare system overload from rapid depressurisation of the compressor train. The anti-surge control system — typically a dedicated Compressor Controls Corporation (CCC) Series 5+ controller or Siemens Dresser-Rand DATUM controller — continuously monitors the compressor operating point (flow vs. head or dP/P) relative to the surge control line and opens the recycle valve (anti-surge bypass valve) to increase compressor inlet flow and move the operating point away from surge. AI systems process rendered compressor performance map display images — real-time plots of compressor operating point against surge control line and surge line on the compressor map, rendered from the anti-surge controller DCS display — to classify surge proximity: normal (operating point right of surge control line), approaching surge control line (recycle valve opening required), crossing surge control line (automatic surge control actuation required), and full surge (emergency compressor trip required).

An adversarial perturbation on a rendered compressor performance map display image that suppresses the operating point’s approach to the surge control line — applying a ±10 DN rightward shift to the pixel position encoding the operating point marker on the compressor map (moving the apparent operating point from the left edge of the normal operating region toward the centre of the map, away from the surge control line) — causes the AI to classify a compressor approaching surge as operating at safe normal operating margin, suppressing the recycle valve opening that the anti-surge control action requires. With the recycle valve not opened, the compressor enters surge: the cyclic flow reversal produces shaft vibration peaks that exceed the automatic trip threshold within 2–5 surge cycles (depending on compressor size and surge severity). If the automatic trip initiates normally, the consequence is a production trip and the potential for seal damage requiring maintenance. If the dry gas seals are damaged by surge before the trip, high-pressure associated gas (containing H₂S in sour gas service) is released into the compressor hall at compressor discharge pressure — an immediate hydrocarbon explosion hazard in the enclosed FPSO module. NORSOK S-001 Section 6.3 requires gas compressors in offshore service to have anti-surge protection systems designed to prevent surge at all operating conditions — adversarial suppression of the anti-surge AI adds a classification failure layer above the anti-surge controller hardware.

2. Wet-gas separator and interstage scrubber knock-out drum level display AI (Emerson Fisher FIELDVUE AI, ABB Totalflow AI, Endress+Hauser Pilomax WAP level AI — FPSO compressor inlet scrubber level display AI)

Liquid carryover to a gas compressor — the entrainment of liquid droplets or slugs of liquid hydrocarbons or water from the wet-gas separator or interstage scrubber knock-out (KO) drum into the compressor inlet — is the most common cause of catastrophic compressor mechanical failure in offshore gas compression service. At FPSO operation with high-GOR (gas-oil ratio) or high-water-cut wells, the gas stream entering the compressor train carries entrained liquid droplets (condensate, crude oil, or formation water) at flow rates that depend on the separator KO drum level and gas flow velocity. If the KO drum liquid level rises above the high-level trip setpoint — due to increased liquid production from the wells, separator performance degradation, or KO drum drain valve malfunction — liquid overflows the KO drum inlet mist extractor and enters the gas stream at rates that can produce a liquid slug of 0.1–10 litres per second at the compressor inlet. A liquid slug entering a centrifugal compressor at full operating speed (10,000–15,000 RPM) produces a hydraulic impulse load on the impeller that can shear impeller vanes, fracture the impeller hub, and destroy the compressor in a single event — the “liquid ingestion” failure mode documented in multiple FPSO gas compressor incidents (Nexen Long Lake 2011, Shell Prelude FLNG 2019 production trip). AI systems process rendered KO drum level display images — strip chart renders or bar chart displays of the scrubber liquid level from differential pressure level transmitters or guided wave radar level gauges — to classify liquid carryover risk: normal (liquid level below 60% of scrubber volume), elevated (level above 60%, drain rate increase required), high (level above 80%, compressor inlet flow reduction required), and emergency (level at high-high trip setpoint, compressor trip required).

An adversarial perturbation on a rendered KO drum level display image that suppresses a rising liquid level — applying a ±8 DN downward shift to the pixel region encoding the level indicator above the elevated alarm threshold (reducing the apparent level from the elevated or high zone to the normal operating range) — causes the compressor monitoring AI to classify an approaching liquid-high level as normal KO drum operation, suppressing the drain rate increase and compressor inlet flow reduction that the elevated level requires. As the undetected liquid level continues to rise and reaches the high-high trip setpoint, the compressor trips on liquid high-level — but not before the liquid level has overflowed the KO drum mist extractor, potentially introducing a liquid slug to the compressor inlet during the seconds before the compressor rundown occurs. At FPSO gas compression pressures (100–300 bar discharge), even a small liquid slug entering the compressor during the rundown period after trip can cause impeller damage. NORSOK S-001 Section 6.3.4 requires that compressor protection systems include high-liquid-level trips on inlet scrubbers — adversarial suppression of the KO drum level AI delays recognition of the approaching trip condition, reducing the operator response window available for controlled compressor unloading before the automatic trip occurs.

3. Emergency shutdown (ESD) valve position display AI (Rotork SKILMATIC AI, Biffi TIMES4 AI, Flowserve Limitorque AI — FPSO ESD valve position and diagnostic display AI)

Emergency shutdown (ESD) systems on offshore FPSOs — networks of fail-safe closed isolation valves (typically ball valves or gate valves with pneumatic or hydraulic actuators and spring-return closure on loss of supply pressure) that isolate hydrocarbon inventories in segments of the process train on Emergency Shutdown Device (ESD) actuation — are the primary layer of protection against escalating offshore hydrocarbon releases. NORSOK S-001 requires ESD systems to be designed, installed, and tested to ensure that each ESD valve achieves positive isolation of the associated hydrocarbon inventory within the specified closure time (typically 10–30 seconds for process isolation valves; 5–10 seconds for riser isolation valves at the FPSO hull) on ESD signal. The ESD valve position monitoring system — open/close position feedback from each ESD valve using limit switches, positioner feedback, or solenoid current monitoring — provides continuous indication of ESD valve status on the DCS ESD mimic display and is the primary source of information for Operations teams to verify that the ESD system has achieved isolation following ESD actuation. AI systems process rendered ESD valve position display images — schematic mimic images of the FPSO process flow diagram with ESD valve position overlays (green = open, red = closed, yellow = intermediate, grey = fault) — to classify ESD system readiness and post-ESD isolation status: all valves in ready state (ESD system ready for actuation), valve position fault (one or more valves indicating intermediate or fault state — maintenance required), ESD isolation achieved (all valves in required state post-ESD), and isolation failure (one or more valves failed to achieve required position post-ESD — manual isolation required).

An adversarial perturbation on a rendered ESD valve position display image that suppresses an ESD valve fault indication — applying a ±8 DN colour shift to the pixel region encoding a yellow (intermediate) or grey (fault) valve position indicator (shifting the apparent valve colour to the green open or red closed state expected for that valve in the ESD system state) — causes the ESD monitoring AI to classify a faulted or partially actuated ESD valve as correctly positioned, suppressing the maintenance action and manual isolation backup that the valve fault classification requires. The Piper Alpha disaster was directly caused by the failure of the ESD system to isolate the condensate injection system from the fire in Module C: the ESD valves on the gas risers from the connected platforms (Tartan, Claymore, MCP-01) remained open because the platform operators on those platforms continued pumping hydrocarbon to Piper Alpha even after the initial explosion, as the Piper Alpha operations team could not communicate the need to shut down the gas supply. In a modern AI-monitored ESD system, adversarial suppression of an ESD valve fault indication prevents the Operations team from identifying that the ESD isolation has not been fully achieved — leaving a hydrocarbon source connected to an active fire or explosion scenario exactly as occurred at Piper Alpha. BSEE SEMS 30 CFR Part 250 Subpart S requires SEMS programs that include ESD system testing and maintenance — but does not address adversarial robustness of AI systems classifying rendered ESD valve position display images.

4. Flare knock-out (KO) drum liquid level display AI (Emerson Rosemount flare KO drum AI, Fisher Fieldvue flare level AI, ABB LevelMaster flare drum AI — FPSO flare system KO drum level AI)

The flare system on an FPSO — the elevated flare stack and associated piping, knock-out drums, and liquid seal pots that receive hydrocarbon vapours from pressure relief valves (PRVs), ESD depressurisation, and normal production venting — is a critical safety system that must be available to accept the maximum emergency vapour release without causing liquid carryover to the flare tip. Liquid carryover to the flare — the entrainment of liquid hydrocarbons from an overfilled flare KO drum into the flare vapour stream — produces burning liquid droplets ejected from the flare tip at high velocity, creating a rain of burning liquid that can fall on FPSO topsides equipment, personnel, and the FPSO hull. The flare KO drum — a gravity separator vessel (typically 2–4 metres diameter, 10–15 metres long, horizontal orientation) that removes entrained liquid droplets from the flare vapour stream by residence time settling — has a maximum liquid level above which liquid begins to overflow into the vapour outlet and reach the flare tip. The flare KO drum liquid level rises during large ESD depressurisation events (such as a platform-wide ESD following a gas release) when the sudden pressure relief of all process segments simultaneously sends large liquid volumes to the flare header. AI systems process rendered flare KO drum level display images — bar chart or strip chart displays of the KO drum liquid level from differential pressure or guided wave radar level transmitters — to classify flare system liquid loading: normal (level below 60% of KO drum volume), elevated (level above 60%, flare liquid pump-out rate increase required), high (level above 80%, production ESD limitation required), and emergency (level approaching overflow — immediate flare depressurisation load reduction required, potential for liquid carryover to flare tip).

An adversarial perturbation on a rendered flare KO drum level display image that suppresses a rising liquid level — applying a ±10 DN downward shift to the pixel region encoding the KO drum level indicator approaching the high-level warning (reducing the apparent level from the elevated zone to the normal mid-drum operating range) — causes the flare system AI to classify an approaching KO drum overflow as normal flare system loading, suppressing the pump-out rate increase and production ESD load limitation that the high level requires. If the KO drum subsequently overflows during a concurrent large ESD depressurisation event — such as a gas release that triggers platform-wide ESD — liquid carryover to the flare tip produces the burning liquid rain scenario: ignited droplets of crude oil or condensate falling from a flare tip 30–60 metres above the FPSO deck can land on topsides equipment, process modules, and personnel muster areas. NORSOK S-001 Section 9 requires that flare systems be designed for the maximum credible simultaneous release rate, with KO drums sized to prevent liquid carryover under all design basis release scenarios — adversarial suppression of the flare KO drum level AI reduces the available response margin before the design basis is exceeded by preventing early detection of abnormal KO drum liquid loading.

Integration: offshore FPSO gas compression AI scanning with Glyphward pre-scan gate

The Glyphward scan gate for offshore FPSO gas compression AI belongs at every rendered-image ingestion boundary in the FPSO gas compression safety AI pipeline — before compressor anti-surge control display AI processes rendered compressor performance map images, before wet-gas scrubber KO drum level AI processes rendered level display images, before ESD valve position display AI processes rendered mimic schematic images, and before flare KO drum level AI processes rendered level display images. Threshold 30 reflects the Piper Alpha consequence anchor (167 killed: ESD failure to isolate condensate system from fire) and the enclosed FPSO topsides environment with no land evacuation option (helicopter evacuation only, limited by weather).

import asyncio, base64, hashlib, json
from datetime import datetime, timezone
from enum import Enum
from pathlib import Path

import httpx

GLYPHWARD_API_KEY = "YOUR_GLYPHWARD_API_KEY"
GLYPHWARD_SCAN_URL = "https://glyphward.com/v1/scan"

# Offshore FPSO gas compression AI contexts: threshold 30
# NORSOK S-001:2023 Technical Safety (offshore process);
# BSEE SEMS 30 CFR Part 250 Subpart S (US Gulf of Mexico);
# ISO 13702:2015 Control and Mitigation of Fires and Explosions Offshore.
FPSO_GAS_COMPRESSION_THRESHOLD = 30


class FPSOGasCompressionAIContext(Enum):
    COMPRESSOR_SURGE     = "compressor_surge"     # Anti-surge control display AI
    KO_DRUM_LEVEL        = "ko_drum_level"         # Wet-gas scrubber KO drum level AI
    ESD_VALVE_POSITION   = "esd_valve_position"   # ESD valve position display AI
    FLARE_KO_DRUM_LEVEL  = "flare_ko_drum_level"  # Flare system KO drum level AI


class AdversarialFPSOGasImageError(Exception):
    """Raised when Glyphward detects adversarial content in an offshore FPSO
    gas compression AI rendered display image above threshold 30.

    Consequence if not raised:
    - COMPRESSOR_SURGE: surge approach suppressed → uncontrolled surge →
      DGS seal failure → high-pressure associated gas release in compressor
      hall → hydrocarbon explosion in enclosed FPSO module.
    - KO_DRUM_LEVEL: liquid carryover risk suppressed → liquid slug into
      compressor → impeller fracture at 10,000-15,000 RPM → catastrophic
      mechanical failure + high-pressure gas release.
    - ESD_VALVE_POSITION: ESD valve fault suppressed → incomplete isolation
      after ESD actuation → hydrocarbon source remains connected to fire;
      Piper Alpha 1988 mechanism (167 killed, gas riser not isolated).
    - FLARE_KO_DRUM_LEVEL: KO drum overflow suppressed → liquid carryover
      to flare tip → burning liquid rain on FPSO topsides + personnel.
    Fail-safe: halt AI classification; initiate manual compressor walkdown
    and ESD valve position verification per NORSOK S-001 Section 6.3;
    contact Offshore Installation Manager (OIM) before resuming
    AI-driven FPSO gas compression safety advisory functions.
    """

    def __init__(self, scan_id: str, score: int,
                 context: FPSOGasCompressionAIContext,
                 fpso_id: str, module_id: str,
                 flagged_region: dict | None = None) -> None:
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.fpso_id = fpso_id
        self.module_id = module_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial FPSO gas compression image: "
            f"context={context.value} score={score} "
            f"fpso={fpso_id} module={module_id} scan_id={scan_id}"
        )


async def scan_fpso_gas_image(
    image_bytes: bytes,
    context: FPSOGasCompressionAIContext,
    fpso_id: str,
    module_id: str,
    client: httpx.AsyncClient,
) -> dict:
    """Scan an offshore FPSO gas compression AI rendered display image.

    Fail-safe contract: AdversarialFPSOGasImageError or httpx error →
    halt FPSO gas compression AI classification for the affected monitoring
    zone; require manual compressor walkdown (NORSOK S-001 Section 6.3)
    and ESD valve position physical verification before resuming
    AI-driven compressor safety or ESD advisory functions.
    """
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"fpso_gas:{context.value}:{fpso_id}:{module_id}",
        "metadata": {
            "fpso_id": fpso_id,
            "module_id": module_id,
            "context": context.value,
            "image_sha256": image_hash,
        },
    }
    resp = await client.post(
        GLYPHWARD_SCAN_URL,
        headers={"Authorization": f"Bearer {GLYPHWARD_API_KEY}"},
        json=payload,
        timeout=4.0,
    )
    resp.raise_for_status()
    result = resp.json()

    if result["score"] > FPSO_GAS_COMPRESSION_THRESHOLD:
        raise AdversarialFPSOGasImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            fpso_id=fpso_id,
            module_id=module_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_fpso_gas_image at each FPSO gas compression AI rendered-image ingestion boundary: before compressor anti-surge control map display AI (threshold 30), before wet-gas scrubber KO drum level display AI (threshold 30), before ESD valve position mimic display AI (threshold 30), and before flare KO drum level display AI (threshold 30). On AdversarialFPSOGasImageError for ESD_VALVE_POSITION context: immediately notify the Offshore Installation Manager (OIM) and conduct physical walkdown to verify ESD valve positions against the required state — do not rely on display readback for isolation verification until after manual position confirmation per NORSOK S-001 ESD testing and maintenance procedure. See also: LNG liquefaction cold box AI prompt injection (related cryogenic gas compression AI adversarial context) and pipeline integrity inspection AI prompt injection (related offshore pipeline AI adversarial context). Get early access

Related questions